Queue Prediction

Removing EWMA greatly increased time to target accuracy and worsened normalized expected delay in ablation experiments. Moderate queue-predictor smoothing performed better than overly slow or overly reactive estimates in the reported experiments. FedQueue's theory is not tied to EWMA specifically and requires sub-Gaussian queue-prediction errors. FedQueue estimates queue delays with an exponentially weighted moving average updated when client updates arrive. Online queue estimates allow FedQueue to reduce local work during predicted long queues and increase it when admission is expected to be faster.